Every operator we sit down with has a list of AI projects somebody on the team is excited about. The list is rarely wrong. Almost everything on it is worth doing eventually. The problem is that the list is also longer than the team's capacity to execute, and trying to do it all at once is the single most common reason AI initiatives stall inside a mid-market business.
The interesting question is not which projects on the list are good. They are mostly all good. The interesting question is which to do this quarter, which to do next, and which to leave on the list until something changes. That is the prioritization problem, and most companies do not have a clean framework for it.
Why the obvious approaches fail
Most leadership teams default to one of two approaches. The first is the impact ranking. Estimate the dollar value of each project, sort by size, do the biggest one first. The second is the enthusiasm ranking. Whichever project the most senior person in the room is most excited about, do that one.
Both fail for the same reason. They optimise for the wrong variable. The biggest project is usually the slowest, the riskiest, and the one most likely to consume the political capital of the team before delivering anything visible. The most exciting project is usually the one with the weakest ROI and the longest path to production. The right sequencing is not about size or excitement. It is about momentum.
The framework we use
Every candidate project goes into one of four buckets. The buckets are scored against three variables: time to first value, confidence in the outcome, and political surface area. The math is simple. The discipline is in actually doing it.
Bucket 1: Recovery work
These are projects that take time the business is already losing and give it back. Automating invoice intake. Drafting first-pass replies to common customer questions. Generating weekly reports the finance team is currently writing by hand. The dollar impact is moderate. The time to value is fast. The confidence is high. The political surface area is small, because nobody is attached to the work being recovered.
Almost every mid-market company should be doing recovery work first. It funds the next round of work by freeing up the team that has to execute it, and it builds the operational muscle for shipping AI inside the business. The visible early win is more valuable than its dollar value suggests, because it changes the conversation about whether AI is real inside the company.
Bucket 2: Defensible advantage
These are projects that, if they work, change the company's position in the market. AI-augmented underwriting that is faster than the competition. A pricing system that adjusts to demand in ways the competitors cannot match. A customer experience that runs on data the company uniquely has access to. The dollar impact is large. The time to value is longer. The confidence is moderate, because the work touches the parts of the business that are hard to change.
These should be the second round of work, not the first. The recovery work earns the right to attempt them by proving that the team can ship AI projects to production, and by freeing up the senior time the harder work will require. Starting with these directly, with no recovery work underneath, is the failure pattern we see most often.
Bucket 3: Optionality bets
These are projects where the value is uncertain but the cost of finding out is low. A small experiment with a new model. A pilot of an emerging vendor. A workflow that might work if a piece of the underlying technology improves in the next six months. The right approach is to fund a few of these in parallel, each with a hard time-box, and to kill the ones that have not produced a signal by the time-box.
Most companies under-invest in optionality bets, because the dollar return is hard to forecast and the team building them feels exposed. The under-investment is a mistake. The cost of an option that does not pay off is small. The cost of missing the option that does is large, and the gap between the two is the difference between being two years ahead and two years behind on the next wave.
Bucket 4: Wait-and-see
These are projects that are good ideas in principle but cannot be done well today, because the underlying technology, the data layer, or the organisational readiness is not yet there. The temptation is to start them anyway. The discipline is to write them down, agree on what would need to be true for them to move into another bucket, and revisit them quarterly.
Most leadership teams want a project moved into wait-and-see roughly once a quarter. Doing this honestly is a sign of a healthy AI strategy, not a failed one. The companies that get themselves into trouble are the ones that promote every interesting idea into the active list, then quietly drop them six months later when the team is exhausted.
How the buckets should be funded
A first-year AI budget for a mid-market company usually wants to split roughly sixty, twenty-five, fifteen, and zero. Sixty percent into recovery work, because that is where momentum and credibility are built. Twenty-five percent into one defensible-advantage project, ideally one that depends on the team that the recovery work has already freed up. Fifteen percent split across two or three optionality bets, each with a clear time-box. Zero percent into wait-and-see, which is a list, not a project.
By the second year, the split shifts. Less recovery work, because the easy gains have been taken. More defensible-advantage work, because the team is now ready. The optionality bets start to graduate into the other buckets. The split is not the same forever. The framework is.
How we help
The framework is straightforward. The work is in applying it honestly. Most leadership teams cannot sort their own list cleanly, because every project on it belongs to someone on the team, and every owner is going to argue for their project being in a more aggressive bucket than it should be. The sort works better when an outside operator is the one doing it, because the outsider has no allegiance to any of the projects and is being paid specifically for the call.
That is the version of strategy work we do most often. Not deciding what the company should build. Deciding what it should build first, what it should build next, and what it should put down for now. The sort is the engagement. Everything else follows from it.
